SolsticeAI/Nanbeige4.3-33B-MLX-UltraOptimised-Curated-Protium

Nanbeige4.3-33B (Apple Silicon MLX UltraOptimised • Curated • Protium-1.0)

Official Flagship MLX Suite • 33.1B Dense Teacher Model • 3-Bit to FP16 • Protium-1.0 Dynamic Reasoning • Apple Silicon Unified Memory

Original Teacher Model by Nanbeige / BOSS Zhipin • Native MLX Conversions & Protium-1.0 Cognitive Architecture by Solstice-AI

---

Executive Summary

SolsticeAI/Nanbeige4.3-33B-MLX-UltraOptimised-Curated-Protium is the official Solstice-AI consolidated Apple Silicon MLX release of Nanbeige4.3-33B, the open-weights 33.1B dense teacher model developed by Nanbeige LLM Lab (BOSS Zhipin / Kanzhun Limited).

Engineered specifically for high-throughput inference on Apple M-series chips (M1/M2/M3/M4 Pro, Max, and Ultra), this release consolidates the entire MLX quantization spectrum from 3-bit to full unquantized FP16 into a single unified repository.

---

Directory Catalog & Hardware Recommendations

Each subdirectory contains a standalone, complete MLX package (config.json, tokenizer.json, generation_config.json, and quantized safetensors shards):

Subdirectory	Target Bitrate	Weight Footprint	Target Apple Silicon Hardware	Description
4Bit/	4-bit	~18.6 GB	24GB - 36GB Unified Memory (M-series Pro / Max)	⭐️ Recommended Production Driver: Exceptional balance of high token generation speeds (>45 tok/s on Max) and deep reasoning fidelity.
3Bit/	3-bit	~14.2 GB	16GB Unified Memory (Base MacBook Pro / Air)	Ultra-compact memory fit for tight RAM constraints.
5Bit/	5-bit	~22.5 GB	32GB - 48GB Unified Memory	High-precision sweet spot for complex multi-turn coding.
6Bit/	6-bit	~26.2 GB	36GB - 64GB Unified Memory	Near-lossless retention of abstract syntax trees and math proofs.
8Bit/	8-bit	~34.1 GB	48GB - 64GB Unified Memory	Reference 8-bit precision for zero perceptible degradation.
fp16/	16-bit	~66.2 GB	64GB - 128GB Unified Memory (Mac Studio / Max / Ultra)	Uncompressed baseline master weights in pure FP16.

---

Quickstart Guide

1. Installation

pip install mlx-lm

2. Python Inference

Download and load your desired bitrate subfolder (e.g. 4Bit):

from mlx_lm import load, generate

repo_id = "SolsticeAI/Nanbeige4.3-33B-MLX-UltraOptimised-Curated-Protium"
subfolder = "4Bit"

model, tokenizer = load(repo_id, model_path=subfolder)

prompt = "Write a high-performance concurrent worker pool in Go with graceful context cancellation."

if hasattr(tokenizer, "apply_chat_template") and tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=True)
print(response)

---

Benchmark Performance Matrix (~30B Class)

Nanbeige4.3-33B benchmarked under maximum reasoning effort against the frontier ~30B class:

Category	Benchmark	Nanbeige4.3-33B	Qwen3.8-27B	Muse-Glimmer-30B	Gemma4-31B
Agentic Coding	Terminal-Bench 2.1	80.9	73.0	61.8	47.2
	Terminal-Bench 4.0	23.8	6.3	1.6	0.0
	DeepSWE	65.5	42.5	4.4	2.7
	SWE-Bench Pro	60.3	61.7	36.9	51.2
	SWE-Atlas-QnA	50.8	41.9	17.7	3.3
	NL2Repo	53.0	48.3	27.4	33.8
Agentic Cowork	Automation Bench	54.2	38.3	15.2	9.7
	JobBench	44.8	34.5	16.6	9.0
	GDPval	84.8	86.4	75.8	71.4
	Claw-Eval (pass^3)	78.3	75.2	57.3	44.2
	Toolathlon-Verified	70.4	68.5	47.2	27.8
	DeepPlanning	63.3	49.2	21.4	19.4
Reasoning	GPQA Diamond	89.5	89.2	83.5	91.3
	AIME 2026	95.2	94.5	94.7	89.2
	AA-LCR	84.0	82.0	83.3	72.0

---

Credits & Upstream Attributions

• Original Foundation Model: Nanbeige / BOSS Zhipin (Kanzhun Limited) for training and releasing the open-weights 33.1B dense foundation model under the Apache 2.0 license.
• Apple Silicon MLX Suite: Quantized, converted, and verified natively by Solstice-AI using mlx-lm.
• Cognitive Architecture: Protium-1.0 Objective-Gated Dynamic Reasoning Engine developed by Solstice-AI.

---

License & Responsible Use

This model card and quantization suite inherit the Apache 2.0 License from upstream Nanbeige. Free for commercial and research use.

---

Citation

@misc{nanbeige43mlxprotium2026,
  title        = {Nanbeige4.3-33B-MLX-UltraOptimised-Curated-Protium},
  author       = {Solstice-AI Team and Nanbeige Team},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/SolsticeAI/Nanbeige4.3-33B-MLX-UltraOptimised-Curated-Protium}}
}

Solstice-AI • Sovereign High-Efficiency Machine Intelligence